All work

06 / 08AI System2026Prototype

Sanjivani AI

Multimodal crisis intelligence for Bihar floods

Reads distress posts, satellite imagery and supply history in one pipeline — triage from text, flood extent from pixels, resource forecasts per district.

Source
Context
Flood response · Bihar
Modalities
Text · Satellite · Tabular
Status
Prototype
Year
2026

01 Problem

In a Bihar flood, the signals that matter — a stranded family’s post, a satellite pass over a river, last season’s supply numbers — arrive in different formats to different people. Sanjivani puts them in one pipeline so a relief planner sees one picture per district.

02 Approach

  1. Triage the text

    A DistilBERT classifier tags posts for urgency, resource needs and vulnerability; a location extractor maps them to districts.

  2. Read the imagery

    U-Net (ResNet50 encoder) segments flood extent, YOLOv8 detects objects, and a change-detection step compares passes.

  3. Forecast the need

    XGBoost models, one per resource type, predict requirements per district from engineered features.

  4. Serve it

    A FastAPI backend (analyze-tweet, analyze-image, forecast per district) with PostGIS storage and a multi-page Streamlit dashboard with maps.

  5. Ship it

    Docker Compose for development and production, with a pytest suite over the API, NLP and helpers.

03 Results

Modalities: text · imagery · tabular
3
API endpoints
5
Tests passing
34/34
  • NLP, vision and forecasting modules trained end-to-end and served behind one API; 34/34 tests passing.
  • So far trained on synthetic data (350 posts, 200 satellite images), so the scores are not meaningful yet — retraining on real social and Sentinel-2 data is the next step.

Honest status: prototype. Metrics will be published once measured on real data.

04 Stack

  • DistilBERT
  • U-Net
  • YOLOv8
  • XGBoost
  • FastAPI
  • Streamlit
  • PostGIS
  • Docker
Next case study · 07Agentic AI Core